scarecrow0 / bgnn-sgg Goto Github PK

Hi, Rongjie,
Thanks for your job. It is another great attempt to solve the unbiased issue in scene graph generation after the paper "Unbiased Scene Graph Generation from Biased Training". I wonder if you have refactored the code and when it will be available.

I am looking forward to your reply, and thanks again.

Thanks for your paper! The model looks very effective in unbiased scene graph generation according to the paper. When will the code be released? I am looking forward to it!

Thank you for the awesome work.

I would like to work on SGG with OI v6, and it seems that your work is, to the best of my knowledge, the first to evaluate on OI v6. I know it takes time to clear codebases before releasing the whole codebase, thus I am wondering is it possible to release the data loader/loading functions first so that people can evaluate SGG models under the same setting?

Moreover, I've a couple of questions and I hope to get you help to clarify:

• I found the visual relationships annotations in OIv6 do not seem to include bounding boxes but only union boxes. In this case, how did you obtain ground truth subject/object boxes?
• It seems only 21 out of 30 predicates (excluding is predicate) shows up in the current OIv6 testing split. How do you calculate mean recall in this case (e.g. simply neglecting them?)

Thank you!

Dear authors,

First of all, thank you for the nice paper! The main idea of refining graph structure using estimated confidence is impressive. I'm looking forward to your source code release! :)

I would like to have you clarify some questions:

• Where does the supervision of the confidence module (RCE), i.e., L_{rce} come from? You mentioned L_{rce} consists of L_m and L_b, what're the meanings of these two terms?
• In Table 1, I feel it's unfair to boldlize your BGNN performance as, simply, your mRs are not superior to that of PCPL [52]; especially, they're using an inferior backbone, i.e., VGG16.
• Performance without graph constraint would be interesting to see & be compared with existing methods, as that is the fairer metric that relaxes the unreasonable constraint of only one relation is being predicted for each pair.
• Figure 4 is not mentioned in the paper. Further explanation and clarification are nice to be included in the main text.
• You mentioned that the hyperparameter ρ ("lo-") is -5, which is not within [0, 1] interval.

*Edit: I found further explanations regarding the supervision of RCE in the Appendix and it's clear now.